This project is a minimized runnable project cut from trojanzoo, which contains more datasets, models, attacks and defenses. This repo will not be maintained.
This is a minimum code implementation of our USENIX'22 paper On the Security Risks of AutoML.
Abstract
The artifact discovers the vulnerability gap between manual models and automl models against various kinds of attacks (adversarial, poison, backdoor, extraction and membership) in image classification domain. It implements all datasets, models, and attacks used in our paper.
We expect the artifact could support the paper's claim that automl models are more vulnerable than manual models against various kinds of attacks, which could be explained by their small gradient variance.
Checklist
- Binary: on pypi with any platform.
- Model: ResNet and other model pretrained weights are available with
--officialflag to download them automatically at first running. - Data set: CIFAR10, CIFAR100 and ImageNet32.
Use--downloadflag to download them automatically at first running.
ImageNet32 requires manual set-up at their website due to legality. - Run-time environment:
At any platform (Windows and Ubuntu tested).
Pytorchandtorchvisionrequired. (CUDA recommended)
adversarial-robustness-toolboxrequired for extraction attack and membership attack. - Hardware: GPU with CUDA support is recommended.
- Execution: Model training and backdoor attack would be time-consuming. It would cost more than half day on a Nvidia Quodro RTX6000.
- Metrics: Model accuracy, attack success rate, clean accuracy drop, cross entropy, f1 score, and auc.
- Output: console output and saved model files (.pth).
- Experiments: OS scripts.
- How much disk space is required (approximately):
less than 5GB. - How much time is needed to prepare workflow (approximately): within 1 hour.
- How much time is needed to complete experiments (approximately): 3-4 days.
- Publicly available: on GitHub.
- Code licenses: GPL-3.
- Archived: GitHub commit #XXXXXXX (todo).
Description
How to access
- GitHub
pip install -e . - PYPI
pip install autovul - Docker Hub
docker pull local0state/autovul - GitHub Packages
docker pull ghcr.io/ain-soph/autovul
Hardware Dependencies
Recommend to use GPU with CUDA and CUDNN.
Less than 5GB disk space is needed.
Software Dependencies
You need to install python==3.9, pytorch==1.9.x, torchvision==0.10.x manually.
ART (IBM) required for extraction attack and membership attack.
pip install adversarial-robustness-toolbox
Data set
CIFAR10, CIFAR100 and ImageNet32.
Use --download flag to download them automatically at first running.
ImageNet32 requires manual set-up at their website due to legality.
Models
ResNet and other model pretrained weights are available with --official flag to download them automatically at first running.
Installation
- GitHub
- PYPI
pip install autovul - Docker Hub
- GitHub Packages
(optional) Config Path
You can set the config files to customize data storage location and many other default settings. View /configs_example as an example config setting.
We support 3 configs (priority ascend):
- package:
(DO NOT MODIFY)
autovul/base/configs/*.yml
autovul/vision/configs/*.yml - user:
~/.autovul/configs/base/*.yml
~/.autovul/configs/vision/*.yml - workspace:
./configs/base/*.yml
./configs/vision/*.yml
Experiment Workflow
Bash Files
Check the bash files under /bash to reproduce our paper results.
Download Datasets
If you run it for the first time, please run bash ./bash/train.sh "--download" to download the dataset.
Train Models
You need to first run /bash/train.sh to get pretrained models.
Run Attacks
/bash/adv_attack.sh
/bash/poison.sh
/bash/backdoor.sh
/bash/extraction.sh
/bash/membership.sh
Run Other Exps
/bash/grad_var.sh
/bash/mitigation_backdoor.sh
/bash/mitigation_extraction.sh
For mitigation experiments, the architecture names in our paper map to:
- darts-i : diy_deep
- darts-ii : diy_no_skip
- darts-iii: diy_deep_noskip
These are the 3 options for --model_arch {arch} (with --model darts)
Evaluation and Expected Result
Our paper claims that automl models are more vulnerable than manual models against various kinds of attacks, which could be explained by low gradient variance. Therefore, for each attack, we expect automl models to have:
Train
Most models around 96%-97% accuracy on CIFAR10.
Attack
For automl models on CIFAR10,
- adversarial
higher success rate (around 10%). - poison
lower accuracy drop (around 5%). - backdoor
higher success rate (around 2%) lower accuracy drop (around 1%). - extraction
lower inference cross entropy (around 0.3). - membership
higher auc (around 0.04).
Others
- gradient variance
automl with lower gradient variance (around 2.2). - mitigation architecture
deep architectures (darts-i, darts-iii) have larger cross entropy for extraction attack (around 0.5), and higher accuracy drop for poisoning attack (around 7%).
Experiment Customization
Use -h or --help flag for example python files to check available arguments.
893 Dec 28, 2022
16 Dec 14, 2022
43 Nov 11, 2022
14 Nov 29, 2022
16 Oct 27, 2022
65 Dec 10, 2022
57 Dec 12, 2022
2 Dec 13, 2021
724 Jan 04, 2023
778 Jan 02, 2023
12 Dec 18, 2022
3 Oct 06, 2022
16 Jan 08, 2023
54 Dec 06, 2022
43 Dec 29, 2022
42 Dec 31, 2022
46 Dec 15, 2022
18 Aug 11, 2022
2 Mar 19, 2022
197 Dec 09, 2022